Artificial Neural Networks in Financial Markets Trading

Artificial neural networks are sophisticated computer systems that mimic the working of the human brain. Often referred to as artificial intelligence, neural networks can perform various functions of the human brain and have cognitive abilities, that is, the ability to learn, which most computer programs do not. Deep Blue, the famous IBM computer that played chess against Garry Kasparov, and Deep Insight, an expert system that can recognize profitable trading patterns in the stock market, are examples of neural network systems that can perform functions and also learn as they go along.

Neural networks can be trained to perform such functions as speech and handwriting recognition, credit card fraud protection, establishing credit risk limits, loan application processing, and analysis of market research data. Neural networks are also employed in monitoring power plant systems, automatic language translation, text-to-speech conversion, bomb detection, prediction of traffic accidents, medical diagnostics and aircraft radars. In fact, neural networks and Kalman filters are used wherever an unknown state of a dynamic system, which is obscured by randomness or noise, has to be estimated.

The application of artificial neural networks to trading the financial markets is an interesting practical use of this technology. Just as traders are able to recognize patterns of arbitrage opportunities based on their trading experience, artificial neural networks can be trained to identify profitable trading opportunities. This technology has a distinctive edge over traders in that they are able to completely isolate human emotions such as greed and fear from the decision-making process.

An experienced trader can easily identify whether a particular market is bid or offered by looking at a computer screen blinking with price updates. However, traders need coffee breaks to escape from the stress of such highly focused work. Artificial neural networks can perform the same task 24/7, and free of emotions, too. One of the commandments of trading is that, If you have a position, forget your emotion. This is easy to remember, but in practice most traders develop emotional attachment to their positions or their models, even after the markets have proven them wrong.

The saying, The markets may not be smart, but they are always right, is another principle taught to all traders. In the book, When Genius Failed: The Rise and Fall of Long Term Capital Management, there is an anecdote about Lawrence Hillibrand, who insisted that his trading models were right even after he had raked up trading losses of a billion dollars. If smart people had infinite capital, then they may eventually turn out to be always right. However, in its absence, trading can be a humbling experience even for the most experienced. Artificial neural networks can provide an alternative to the human trader and make trading easier, more efficient and more profitable.

The construction of artificial neural networks is very similar to that of the human brain. They have multiple processors consisting of simple processing elements (the artificial neurons) that are connected to one another. In the human brain, there are about 100 billion neurons connected to each other by synapses, an electrochemical contact. Each neuron can be connected with up to 200,000 other neurons, though typically each neuron connects with only 1,000 to 10,000 other neurons. Yet, even the most complex artificial neural network does not have so many artificial neurons connected with each other.

In their report titled, Artificial Neural Networks Technology prepared for the Rome Laboratory, Anderson and McNeil articulate how artificial neural networks can reproduce many of the functions of the human brain by simulating the neural architecture in either computer hardware or software.

There are three layers in any artificial neural network, the input layer, the hidden layer and the output layer. The input later corresponds to the dendrites of human neurons, which accept inputs from other neurons. The hidden layer corresponds to the soma, which processes the inputs, and the output layer corresponds to the axon in human neurons, which convert the processed inputs into output. The number of nodes in the input and output layers are fixed while the hidden layer has a variable number of nodes.

In a trading system based on neural networks, the output layer would typically have three nodes - one each for buy, sell or hold signals.

In the human brain, neurons are clustered together in 3-D space. Creating a similar cluster of artificial neural networks is considered an art form because there are numerous ways in which the artificial neurons can be put together. The clustering of human neurons appears to take place without any limitations or restrictions. However, integrated circuits and microprocessors do limit interconnection. Artificial neural networks, thus, are limited by the physical constraints of silicon chips. Training artificial neural networks, therefore, is a key to their successful practical applications. Whats needed is adaptive learning. In adaptive learnings systems, the neural network does not know the correct answer but is able to find it by learning from its errors. Tuevo Kohonen of Helsinki University of Technology was a pioneer in developing adaptive learning artificial neural networks. At IIT-Mumbai and at Citibank, India, Dr Mahendra Mehta developed neural network systems using adaptive Kalman filters.

Using topological ordering, mathematicians and engineers have begun to develop artificial neural networks that mimic the human brain more closely.The key distinguishing feature of these neural networks is their ability to learn and adapt. There are many learning laws that are used to train the neural networks to think. One of these laws called Hebbs Rule was first formulated in 1949 in a book titled The Organization of Behavior.

Other laws such as Hopfield Law and The Delta Rule were subsequently formulated as modifications or variations of Hebbs Rule.Human beings may sometimes not learn from their past mistakes, but artificial neural networks have their previous mistakes permanently stored in their memory and their algorithms enable them to recognise similar situations and avoid making the same mistake again. This is due to the fact that human brains have an emotional component but the artificial neural networks are strictly pragmatic.

According to Deep Insight, funds totalling $250 billion are managed using artificial neural network technology and other artificial intelligence systems. Citibanks Global Analytics Group developed an artificial neural networks based system in Hong Kong and India using adaptive Kalman filters and artificial neural networks under the leadership of Dr Pratap Sondhi and Dr Mahendra Mehta. This system was successfully used to trade foreign exchange in Tokyo and London during the 1990s.

Daiichi Kangyo Bank used neural networks to improve its performance in the Japanese government bond markets by as much as 75%, while Daiwa Securities used a neural networks system to outperform the Nikkei stock index by 20%.JP Morgan also employed a neural network based trading system and California Scientific Software has developed a system called BrainMaker that uses neural networks for financial forecasting and trading. Brain Trader Investments of Cambridge, Massachusetts, also employs a proprietary neural networks system called Brain Trader that was developed by Joseph Zirilli.

Neural networks based systems can be applied to any liquid financial market such as the major stocks, stock index futures, currencies, natural gas, oil, T-bond futures and gold. They are, however, not suitable for illiquid bonds, structured products and exotic OTC traded derivative instruments.

The performance of some neural networks based systems has been excellent. One such neural network system yielded a return of 29% in 2001 compared to a 13% loss for the S&P 500 index during that year. The average return for hedge funds, according to a Bloomberg study, was 12.3% while the average return for fund-of-hedge funds was 8% in 2002. According to Dr Mahendra Mehta, the CEO of Neural Technologies and Software Ltd, a technology firm based in India, their neural trading system yielded an average annual return of 57.5% on dollar-yen trading from 1994 to 2000. This return was achieved using a five times leverage (quite low by hedge fund standards) when dollar-yen tic-by-tic prices from Reuters were used to test the neural network system. The neural trading system yielded an average annual return of more than 50% for currency pairs such as dollar/euro and euro/yen for the period 1994 to 2000. In less liquid currency pairs such as dollar/Swiss franc and British pound/dollar, the average returns were 34.1% and 23.7%, respectively. Of particular significance is the fact that the maximum draw down using a five times leverage was just 17%.Neural networks are adept in making profitable trading decisions in highly liquid markets. Markets with low liquidity or highly structured financial products are not conducive to the use of neural networks in making trading decisions.In sum, model-driven trading strategies based on neural networks are an alternative worth trying. They do not require any traders. And the computer does not throw tantrums, take coffee breaks or vacations or demand any bonuses at the end of a year